MELGene: knowledge-enhanced multimodel ensemble learning for disease-gene association prediction

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Disease-gene prediction (DGP) is crucial for understanding genetic disease factors, aiding diagnosis, treatment, and personalized medicine.
• Existing DGP methods struggle to model complex interactions between diseases, symptoms, genes, and pathways effectively.
• Robust modeling of these intricate biological relationships is key for accurate phenotype and genotype representation in DGP.

Purpose of the Study:

• To introduce MELGene, a knowledge-enhanced multimodel ensemble learning framework designed for accurate disease-gene prediction.
• To leverage knowledge graphs and adaptive ensemble learning for improved DGP accuracy.
• To demonstrate the framework's effectiveness in capturing complex biological interactions for enhanced gene predictions.

Main Methods:

• Developed MELGene, a framework integrating multiple pretrained knowledge inference models via knowledge graphs.
• Implemented Model-aware Importance Learning for dynamic adjustment of individual model contributions.
• Utilized a dynamic ensemble mechanism to generate robust consensus predictions.

Main Results:

• MELGene demonstrated excellent performance in comprehensive experimental comparisons.
• Ablation experiments confirmed the positive contribution of each framework module.
• Case studies on gastric, lung, and liver cancers validated the biological relevance of predictions through network medicine and literature mining.

Conclusions:

• MELGene provides a flexible and effective framework for disease-gene prediction through knowledge enhancement and adaptive ensemble learning.
• The framework shows significant potential for advancing the understanding of disease mechanisms and supporting personalized medicine.
• MELGene's approach offers a powerful tool for decoding complex genetic underpinnings of diseases.